---
language:
- en
license: apache-2.0
pipeline_tag: video-text-to-text
base_model:
- Qwen/Qwen3-VL-2B-Instruct
- Qwen/Qwen3-4B
tags:
- vision-language-model
- long-video-understanding
- multimodal
- pytorch
- custom_code
---

# Tempo-6B-Stage2 (Pre-Long-Context SFT)

This repository contains the **Stage 2 intermediate checkpoint** for Tempo-6B. 

Unlike earlier training stages, **this checkpoint is fully capable of direct inference**. It has completed short-video and image instruction tuning, making it a strong baseline for general multimodal understanding. However, it has **not** undergone our final Stage 3 Long-Context SFT.

### 🚀 When to use this checkpoint?
- **Custom Fine-Tuning:** An ideal starting point if you want to apply your own long-context SFT curriculum or adapt the model to specialized domains.

### 📊 Ablation Performance
To demonstrate its capabilities, here is the performance of this Stage 2 checkpoint compared to our final Tempo-6B model:

| Model Setting | LongVideoBench | MLVU | Video-MME (Overall) | Video-MME (Long) | LVBench |
| :--- | :---: | :---: | :---: | :---: | :---: |
| **Tempo-6B-Stage2** (w/o ATA) | 61.4 | 67.2 | 66.1 | 56.3 | 47.3 |
| **Tempo-6B-Final** (w/o ATA) | 62.8 | 73.5 | 67.0 | 56.2 | 51.1 |
| **Tempo-6B Final** (w/ ATA) | **65.1** | **75.2** | **67.7** | **57.0** | **52.3** |

*Note: For a detailed analysis, please check our [paper's Ablation Study A](https://arxiv.org/pdf/2604.08120) (Progressive Training Curriculum).*

### 🔗 Links
For the final long video understanding performance, please use our final weights:
- **Final Model:** [Vision-CAIR/Tempo-6B](https://huggingface.co/Vision-CAIR/Tempo-6B)
- **GitHub Code:** [FeiElysia/Tempo](https://github.com/FeiElysia/Tempo)
- **Paper:** [Small Vision-Language Models are Smart Compressors for Long Video Understanding](https://huggingface.co/papers/2604.08120)